基于视觉注意机制的智能车辆目标检测方法研究

This project concerns the urgent needs of the safe and comfortable travel which relates to national economy and the people’s livelihood. Aiming to solve the problem of contradiction between the large scale of perception data and the limited computation resources in the field of intelligent vehicle object detection, this project tries to investigate the object detection methods based on visual attention mechanism. To make the visual attention modeling methods adaptive to the complex and dynamic transportation environment, the key research contents include: investigating to learn the robust eye movement prediction model to extract the regions of interest under complex environment rapidly and accurately; investigating the method for modeling the dynamic time-varying visual attention mechanism, to switch adaptively on different detection mechanism for different kinds of objects; investigating the method for modeling visual attention by the cooperative computation of scene lingual description information and image information, to improve the detection accuracy of small scale objects and occluded objects. This project provides ideas for visual attention modeling under complex and dynamic environment, provides tools for improving the speed and accuracy for intelligent vehicle environment perception, and has important significance on promoting the continuous integration between biological vision and machine vision and improving the development of artificial intelligence.

项目以关系国计民生的安全、舒适出行迫切需求为应用背景，针对智能车辆目标检测存在的大规模感知数据与有限的计算资源之间矛盾的问题，研究基于视觉注意机制的目标检测方法。重点突破适应复杂动态交通环境的视觉注意建模方法，具体研究内容包括：研究对复杂动态环境鲁棒的眼动预测模型，实现感兴趣区域的快速、准确提取；研究面向时变任务的自适应视觉注意建模方法，实现不同目标检测机制的自适应切换；研究场景语言描述信息与图像信息协同计算的视觉注意建模方法，实现小尺度目标和受遮挡目标的高精度检测。本项目研究内容为复杂动态环境下的视觉注意建模提供思路，为提高智能车辆环境感知的速度和精度提供重要手段，对促进生物视觉与机器视觉的不断融合和推动人工智能学科的发展具有重要意义。

项目以关系国计民生的安全、舒适出行迫切需求为应用背景，针对智能车辆目标检测存在的大规模感知数据与有限的计算资源之间矛盾的问题，研究基于视觉注意机制的目标检测方法。重点突破适应复杂动态交通环境的视觉注意建模方法，具体研究内容包括：研究对复杂动态环境鲁棒的眼动预测模型，实现感兴趣区域的快速、准确提取；研究面向时变任务的自适应视觉注意建模方法，实现不同目标检测机制的自适应切换；研究场景语言描述信息与图像信息协同计算的视觉注意建模方法，实现小尺度目标和受遮挡目标的高精度检测。本项目研究内容为复杂动态环境下的视觉注意建模提供思路，为提高智能车辆环境感知的速度和精度提供重要手段，对促进生物视觉与机器视觉的不断融合和推动人工智能学科的发展具有重要意义。